diff --git a/.gitmodules b/.gitmodules
index 3ae861fb..c688aa84 100644
--- a/.gitmodules
+++ b/.gitmodules
@@ -1,3 +1,3 @@
-[submodule "tmk_core/protocol/usb_hid/USB_Host_Shield_2.0-git"]
-	path = tmk_core/protocol/usb_hid/USB_Host_Shield_2.0-git
-	url = https://github.com/felis/USB_Host_Shield_2.0.git
+[submodule "tmk_core/protocol/usb_hid/USB_Host_Shield_2.0-tmk"]
+	path = tmk_core/protocol/usb_hid/USB_Host_Shield_2.0-tmk
+	url = https://github.com/tmk/USB_Host_Shield_2.0.git
diff --git a/converter/usb_usb/Makefile.debug b/converter/usb_usb/Makefile.debug
index e229cd78..830995b4 100644
--- a/converter/usb_usb/Makefile.debug
+++ b/converter/usb_usb/Makefile.debug
@@ -1,6 +1,20 @@
 TARGET = usb_usb_debug
-UNIMAP_ENABLE = yes
+
+#UNIMAP_ENABLE = yes
 #KEYMAP_SECTION_ENABLE = yes
-#LUFA_DEBUG = yes
-OPT_DEFS += -DDEBUG_USB_HOST
+
+# LUFA debug print
+# This may prevent USB enumeration and keyboard init
+LUFA_DEBUG = yes
+# Select one of outputs for debug print
+LUFA_DEBUG_UART = yes
+#LUFA_DEBUG_SUART = yes
+
+# USB_Host_Shield_2.0 debug print
+# This may prevent USB enumeration and keyboard init
+#OPT_DEFS += -DDEBUG_USB_HOST
+
+CONSOLE_ENABLE = yes
+MOUSEKEY_ENABLE = no
+EXTRAKEY_ENABLE = no
 include Makefile
diff --git a/converter/usb_usb/config.h b/converter/usb_usb/config.h
index 87da4a45..7b80fc21 100644
--- a/converter/usb_usb/config.h
+++ b/converter/usb_usb/config.h
@@ -36,4 +36,17 @@ along with this program.  If not, see <http://www.gnu.org/licenses/>.
 /* key combination for command */
 #define IS_COMMAND() (keyboard_report->mods == (MOD_BIT(KC_LSHIFT) | MOD_BIT(KC_RSHIFT))) 
 
+
+// Disable power saving in USB suspend loop but remote wakeup is still valid.
+// This allows keep USB::Task() going during suspend without power down time delay.
+//#define NO_SUSPEND_POWER_DOWN
+
+
+// Disable USB startup wait, which can delays starting UHS2 Task() for 350-600ms.
+//#define NO_USB_STARTUP_WAIT_LOOP
+
+// Disable USB suspend loop, which blocks UHS2 Task() while power saving.
+// Note that this also disables power saving and remote wakeup from keyboard completely.
+//#define NO_USB_SUSPEND_LOOP
+
 #endif
diff --git a/converter/usb_usb/usb_usb.cpp b/converter/usb_usb/usb_usb.cpp
index 086c90ae..50d2578e 100644
--- a/converter/usb_usb/usb_usb.cpp
+++ b/converter/usb_usb/usb_usb.cpp
@@ -35,6 +35,10 @@ along with this program.  If not, see <http://www.gnu.org/licenses/>.
 #include "host.h"
 #include "keyboard.h"
 
+#include "hook.h"
+#include "suspend.h"
+#include "lufa.h"
+
 
 /* KEY CODE to Matrix
  *
@@ -71,8 +75,6 @@ static bool matrix_is_mod =false;
  * This supports two cascaded hubs and four keyboards
  */
 USB usb_host;
-USBHub hub1(&usb_host);
-USBHub hub2(&usb_host);
 HIDBoot<USB_HID_PROTOCOL_KEYBOARD>    kbd1(&usb_host);
 HIDBoot<USB_HID_PROTOCOL_KEYBOARD>    kbd2(&usb_host);
 HIDBoot<USB_HID_PROTOCOL_KEYBOARD>    kbd3(&usb_host);
@@ -81,6 +83,8 @@ KBDReportParser kbd_parser1;
 KBDReportParser kbd_parser2;
 KBDReportParser kbd_parser3;
 KBDReportParser kbd_parser4;
+USBHub hub1(&usb_host);
+USBHub hub2(&usb_host);
 
 
 uint8_t matrix_rows(void) { return MATRIX_ROWS; }
@@ -233,3 +237,28 @@ void led_set(uint8_t usb_led)
     if (kbd3.isReady()) kbd3.SetReport(0, 0, 2, 0, 1, &usb_led);
     if (kbd4.isReady()) kbd4.SetReport(0, 0, 2, 0, 1, &usb_led);
 }
+
+// We need to keep doing UHS2 USB::Task() to initialize keyboard
+// even before USB is not configured.
+void hook_usb_startup_wait_loop(void)
+{
+    matrix_scan();
+}
+
+// We need to keep doing UHS2 USB::Task() to initialize keyboard
+// even during USB bus is suspended and remote wakeup is not enabled yet on LUFA side.
+// This situation can happen just after pluging converter into USB port.
+void hook_usb_suspend_loop(void)
+{
+#ifndef LUFA_DEBUG_UART
+    // This corrupts debug print when suspend
+    suspend_power_down();
+#endif
+    if (USB_Device_RemoteWakeupEnabled) {
+        if (suspend_wakeup_condition()) {
+            USB_Device_SendRemoteWakeup();
+        }
+    } else {
+        matrix_scan();
+    }
+}
diff --git a/tmk_core/common.mk b/tmk_core/common.mk
index 14951ec8..06507ecf 100644
--- a/tmk_core/common.mk
+++ b/tmk_core/common.mk
@@ -50,8 +50,12 @@ endif
 ifeq (yes,$(strip $(CONSOLE_ENABLE)))
     OPT_DEFS += -DCONSOLE_ENABLE
 else
-    OPT_DEFS += -DNO_PRINT
-    OPT_DEFS += -DNO_DEBUG
+    # Remove print functions when console is disabled and
+    # no other print method like UART is available
+    ifneq (yes, $(strip $(DEBUG_PRINT_AVAILABLE)))
+	OPT_DEFS += -DNO_PRINT
+	OPT_DEFS += -DNO_DEBUG
+    endif
 endif
 
 ifeq (yes,$(strip $(COMMAND_ENABLE)))
diff --git a/tmk_core/common/uart.c b/tmk_core/common/avr/uart.c
similarity index 66%
rename from tmk_core/common/uart.c
rename to tmk_core/common/avr/uart.c
index c17649b0..0fec705f 100644
--- a/tmk_core/common/uart.c
+++ b/tmk_core/common/avr/uart.c
@@ -1,4 +1,3 @@
-// TODO: Teensy support(ATMega32u4/AT90USB128)
 // Fixed for Arduino Duemilanove ATmega168p by Jun Wako
 /* UART Example for Teensy USB Development Board
  * http://www.pjrc.com/teensy/
@@ -32,9 +31,42 @@
 
 #include "uart.h"
 
+#if defined(__AVR_ATmega168__) || defined(__AVR_ATmega168P__) || defined(__AVR_ATmega328P__)
+#   define UDRn         UDR0
+#   define UBRRn        UBRR0
+#   define UCSRnA       UCSR0A
+#   define UCSRnB       UCSR0B
+#   define UCSRnC       UCSR0C
+#   define U2Xn         U2X0
+#   define RXENn        RXEN0
+#   define TXENn        TXEN0
+#   define RXCIEn       RXCIE0
+#   define UCSZn1       UCSZ01
+#   define UCSZn0       UCSZ00
+#   define UDRIEn       UDRIE0
+#   define UDRE_vect    USART_UDRE_vect
+#   define RX_vect      USART_RX_vect
+#elif defined(__AVR_ATmega32U4__) || defined(__AVR_ATmega32U2__) || defined(__AVR_AT90USB1286__) || defined(__AVR_AT90USB1287__)
+#   define UDRn         UDR1
+#   define UBRRn        UBRR1
+#   define UCSRnA       UCSR1A
+#   define UCSRnB       UCSR1B
+#   define UCSRnC       UCSR1C
+#   define U2Xn         U2X1
+#   define RXENn        RXEN1
+#   define TXENn        TXEN1
+#   define RXCIEn       RXCIE1
+#   define UCSZn1       UCSZ11
+#   define UCSZn0       UCSZ10
+#   define UDRIEn       UDRIE1
+#   define UDRE_vect    USART1_UDRE_vect
+#   define RX_vect      USART1_RX_vect
+#endif
+
+
 // These buffers may be any size from 2 to 256 bytes.
 #define RX_BUFFER_SIZE 64
-#define TX_BUFFER_SIZE 40
+#define TX_BUFFER_SIZE 256
 
 static volatile uint8_t tx_buffer[TX_BUFFER_SIZE];
 static volatile uint8_t tx_buffer_head;
@@ -47,10 +79,10 @@ static volatile uint8_t rx_buffer_tail;
 void uart_init(uint32_t baud)
 {
 	cli();
-	UBRR0 = (F_CPU / 4 / baud - 1) / 2;
-	UCSR0A = (1<<U2X0);
-	UCSR0B = (1<<RXEN0) | (1<<TXEN0) | (1<<RXCIE0);
-	UCSR0C = (1<<UCSZ01) | (1<<UCSZ00);
+	UBRRn = (F_CPU / 4 / baud - 1) / 2;
+	UCSRnA = (1<<U2Xn);
+	UCSRnB = (1<<RXENn) | (1<<TXENn) | (1<<RXCIEn);
+	UCSRnC = (1<<UCSZn1) | (1<<UCSZn0);
 	tx_buffer_head = tx_buffer_tail = 0;
 	rx_buffer_head = rx_buffer_tail = 0;
 	sei();
@@ -63,11 +95,13 @@ void uart_putchar(uint8_t c)
 
 	i = tx_buffer_head + 1;
 	if (i >= TX_BUFFER_SIZE) i = 0;
+	// return immediately to avoid deadlock when interrupt is disabled(called from ISR)
+	if (tx_buffer_tail == i && (SREG & (1<<SREG_I)) == 0) return;
 	while (tx_buffer_tail == i) ; // wait until space in buffer
 	//cli();
 	tx_buffer[i] = c;
 	tx_buffer_head = i;
-	UCSR0B = (1<<RXEN0) | (1<<TXEN0) | (1<<RXCIE0) | (1<<UDRIE0);
+	UCSRnB = (1<<RXENn) | (1<<TXENn) | (1<<RXCIEn) | (1<<UDRIEn);
 	//sei();
 }
 
@@ -98,27 +132,27 @@ uint8_t uart_available(void)
 }
 
 // Transmit Interrupt
-ISR(USART_UDRE_vect)
+ISR(UDRE_vect)
 {
 	uint8_t i;
 
 	if (tx_buffer_head == tx_buffer_tail) {
 		// buffer is empty, disable transmit interrupt
-		UCSR0B = (1<<RXEN0) | (1<<TXEN0) | (1<<RXCIE0);
+		UCSRnB = (1<<RXENn) | (1<<TXENn) | (1<<RXCIEn);
 	} else {
 		i = tx_buffer_tail + 1;
 		if (i >= TX_BUFFER_SIZE) i = 0;
-		UDR0 = tx_buffer[i];
+		UDRn = tx_buffer[i];
 		tx_buffer_tail = i;
 	}
 }
 
 // Receive Interrupt
-ISR(USART_RX_vect)
+ISR(RX_vect)
 {
 	uint8_t c, i;
 
-	c = UDR0;
+	c = UDRn;
 	i = rx_buffer_head + 1;
 	if (i >= RX_BUFFER_SIZE) i = 0;
 	if (i != rx_buffer_tail) {
diff --git a/tmk_core/common/hook.h b/tmk_core/common/hook.h
index 56fe5677..140e2779 100644
--- a/tmk_core/common/hook.h
+++ b/tmk_core/common/hook.h
@@ -21,6 +21,10 @@ along with this program.  If not, see <http://www.gnu.org/licenses/>.
 #include "keyboard.h"
 #include "led.h"
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 /* -------------------------------------
  * Protocol hooks
  * ------------------------------------- */
@@ -47,6 +51,10 @@ void hook_usb_suspend_loop(void);
  * the "normal" indicator LED status by default. */
 void hook_usb_wakeup(void);
 
+/* Called repeatedly until getting to CONFIGURED state */
+/* Default behaviour: do nothing. */
+void hook_usb_startup_wait_loop(void);
+
 
 /* -------------------------------------
  * Keyboard hooks
@@ -76,5 +84,8 @@ void hook_keyboard_leds_change(uint8_t led_status);
 /* Default behaviour: do nothing. */
 void hook_bootmagic(void);
 
+#ifdef __cplusplus
+}
+#endif
 
 #endif /* _HOOKS_H_ */
diff --git a/tmk_core/common/suspend.h b/tmk_core/common/suspend.h
index 80617a82..f0a668cf 100644
--- a/tmk_core/common/suspend.h
+++ b/tmk_core/common/suspend.h
@@ -5,9 +5,17 @@
 #include <stdbool.h>
 
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 void suspend_idle(uint8_t timeout);
 void suspend_power_down(void);
 bool suspend_wakeup_condition(void);
 void suspend_wakeup_init(void);
 
+#ifdef __cplusplus
+}
+#endif
+
 #endif
diff --git a/tmk_core/doc/hook.txt b/tmk_core/doc/hook.txt
index 1689034e..acd77cce 100644
--- a/tmk_core/doc/hook.txt
+++ b/tmk_core/doc/hook.txt
@@ -9,6 +9,8 @@ Hook function                   | Timing
 `hook_early_init(void)`         | Early in the boot process, before the matrix is initialized and before a connection is made with the host. Thus, this hook has access to very few parameters, but it is a good place to define any custom parameters needed by other early processes.
 `hook_late_init(void)`          | Near the end of the boot process, after Boot Magic has run and LEDs have been initialized.
 `hook_bootmagic(void)`          | During the Boot Magic window, after EEPROM and Bootloader checks are made, but before any other built-in Boot Magic checks are made.
+`hook_usb_startup_wait_loop(void)` | Continuously, until the device gets ready and into USB configured state.
+
 `hook_usb_wakeup(void)`         | When the device wakes up from USB suspend state.
 `hook_usb_suspend_entry(void)`  | When the device enters USB suspend state.
 `hook_usb_suspend_loop(void)`   | Continuously, while the device is in USB suspend state. *Default action:* power down and periodically check the matrix, causing wakeup if needed.
diff --git a/tmk_core/protocol/lufa.mk b/tmk_core/protocol/lufa.mk
index 1b529bde..19fc80d4 100644
--- a/tmk_core/protocol/lufa.mk
+++ b/tmk_core/protocol/lufa.mk
@@ -49,7 +49,20 @@ OPT_DEFS += $(LUFA_OPTS)
 # This indicates using LUFA stack
 OPT_DEFS += -DPROTOCOL_LUFA
 
+ifeq (yes,$(strip $(LUFA_DEBUG)))
+    LUFA_OPTS += -DLUFA_DEBUG
+endif
+
 ifeq (yes,$(strip $(LUFA_DEBUG_SUART)))
     SRC += common/avr/suart.S
     LUFA_OPTS += -DLUFA_DEBUG_SUART
+    # Keep print/debug lines when disabling HID console. See common.mk.
+    DEBUG_PRINT_AVAILABLE = yes
+endif
+
+ifeq (yes,$(strip $(LUFA_DEBUG_UART)))
+    SRC += common/avr/uart.c
+    LUFA_OPTS += -DLUFA_DEBUG_UART
+    # Keep print/debug lines when disabling HID console. See common.mk.
+    DEBUG_PRINT_AVAILABLE = yes
 endif
diff --git a/tmk_core/protocol/lufa/lufa.c b/tmk_core/protocol/lufa/lufa.c
index b5d0b50e..9dadd572 100644
--- a/tmk_core/protocol/lufa/lufa.c
+++ b/tmk_core/protocol/lufa/lufa.c
@@ -56,6 +56,10 @@
 #include "avr/suart.h"
 #endif
 
+#ifdef LUFA_DEBUG_UART
+#include "uart.h"
+#endif
+
 #include "matrix.h"
 #include "descriptor.h"
 #include "lufa.h"
@@ -218,7 +222,9 @@ static void console_task(void)
 */
 void EVENT_USB_Device_Connect(void)
 {
+#ifdef LUFA_DEBUG
     print("[C]");
+#endif
     /* For battery powered device */
     if (!USB_IsInitialized) {
         USB_Disable();
@@ -229,7 +235,9 @@ void EVENT_USB_Device_Connect(void)
 
 void EVENT_USB_Device_Disconnect(void)
 {
+#ifdef LUFA_DEBUG
     print("[D]");
+#endif
     /* For battery powered device */
     USB_IsInitialized = false;
 /* TODO: This doesn't work. After several plug in/outs can not be enumerated.
@@ -575,25 +583,22 @@ static void send_consumer(uint16_t data)
 /*******************************************************************************
  * sendchar
  ******************************************************************************/
-#ifdef CONSOLE_ENABLE
 int8_t sendchar(uint8_t c)
 {
     #ifdef LUFA_DEBUG_SUART
     xmit(c);
     #endif
 
-    bool r = console_putc(c);
-    return (r ? 0 : -1);
-}
-#else
-int8_t sendchar(uint8_t c)
-{
-    #ifdef LUFA_DEBUG_SUART
-    xmit(c);
+    #ifdef LUFA_DEBUG_UART
+    uart_putchar(c);
     #endif
+
+    #ifdef CONSOLE_ENABLE
+    console_putc(c);
+    #endif
+
     return 0;
 }
-#endif
 
 
 /*******************************************************************************
@@ -629,11 +634,15 @@ int main(void)
     SUART_OUT_PORT |= (1<<SUART_OUT_BIT);
 #endif
 
+#ifdef LUFA_DEBUG_UART
+    uart_init(115200);
+#endif
+
     // setup sendchar: DO NOT USE print functions before this line
     print_set_sendchar(sendchar);
     host_set_driver(&lufa_driver);
 
-    print("Keyboard init.\n");
+    print("\n\nKeyboard init.\n");
     hook_early_init();
     keyboard_setup();
     setup_usb();
@@ -645,6 +654,7 @@ int main(void)
 
     keyboard_init();
 
+#ifndef NO_USB_STARTUP_WAIT_LOOP
     /* wait for USB startup */
     while (USB_DeviceState != DEVICE_STATE_Configured) {
 #if defined(INTERRUPT_CONTROL_ENDPOINT)
@@ -652,19 +662,20 @@ int main(void)
 #else
         USB_USBTask();
 #endif
-        matrix_scan();
+        hook_usb_startup_wait_loop();
     }
+    print("\nUSB configured.\n");
+#endif
 
     hook_late_init();
 
     print("\nKeyboard start.\n");
     while (1) {
+#ifndef NO_USB_SUSPEND_LOOP
         while (USB_DeviceState == DEVICE_STATE_Suspended) {
-#ifdef LUFA_DEBUG
-            print("[s]");
-#endif
             hook_usb_suspend_loop();
         }
+#endif
 
         keyboard_task();
 
@@ -690,12 +701,12 @@ static uint8_t _led_stats = 0;
  __attribute__((weak))
 void hook_usb_suspend_entry(void)
 {
-    // Turn LED off to save power
-    // Set 0 with putting aside status before suspend and restore
-    // it after wakeup, then LED is updated at keyboard_task() in main loop
+    // Turn off LED to save power and keep its status to resotre it later.
+    // LED status will be updated by keyboard_task() in main loop hopefully.
     _led_stats = keyboard_led_stats;
     keyboard_led_stats = 0;
-    led_set(keyboard_led_stats);
+
+    // Calling long task here can prevent USB state transition
 
     matrix_clear();
     clear_keyboard();
@@ -707,7 +718,10 @@ void hook_usb_suspend_entry(void)
 __attribute__((weak))
 void hook_usb_suspend_loop(void)
 {
+#ifndef LUFA_DEBUG_UART
+    // This corrupts debug print when suspend
     suspend_power_down();
+#endif
     if (USB_Device_RemoteWakeupEnabled && suspend_wakeup_condition()) {
         USB_Device_SendRemoteWakeup();
     }
@@ -721,10 +735,11 @@ void hook_usb_wakeup(void)
     sleep_led_disable();
 #endif
 
-    // Restore LED status
-    // BIOS/grub won't recognize/enumerate if led_set() takes long(around 40ms?)
-    // Converters fall into the case and miss wakeup event(timeout to reply?) in the end.
-    //led_set(host_keyboard_leds());
-    // Instead, restore stats and update at keyboard_task() in main loop
+    // Restore LED status and update at keyboard_task() in main loop
     keyboard_led_stats = _led_stats;
+
+    // Calling long task here can prevent USB state transition
 }
+
+__attribute__((weak))
+void hook_usb_startup_wait_loop(void) {}
diff --git a/tmk_core/protocol/usb_hid.mk b/tmk_core/protocol/usb_hid.mk
index 495c4bee..bba4292f 100644
--- a/tmk_core/protocol/usb_hid.mk
+++ b/tmk_core/protocol/usb_hid.mk
@@ -4,7 +4,7 @@ USB_HID_DIR = protocol/usb_hid
 #
 # USB Host Shield
 #
-USB_HOST_SHIELD_DIR = $(USB_HID_DIR)/USB_Host_Shield_2.0-git
+USB_HOST_SHIELD_DIR = $(USB_HID_DIR)/USB_Host_Shield_2.0-tmk
 USB_HOST_SHIELD_SRC = \
 	$(USB_HOST_SHIELD_DIR)/Usb.cpp \
 	$(USB_HOST_SHIELD_DIR)/usbhid.cpp \
diff --git a/tmk_core/protocol/usb_hid/USB_Host_Shield_2.0-git b/tmk_core/protocol/usb_hid/USB_Host_Shield_2.0-git
deleted file mode 160000
index ed08df7e..00000000
--- a/tmk_core/protocol/usb_hid/USB_Host_Shield_2.0-git
+++ /dev/null
@@ -1 +0,0 @@
-Subproject commit ed08df7e68af06a5612d938b1bbf55cc4458518a
diff --git a/tmk_core/protocol/usb_hid/USB_Host_Shield_2.0-tmk b/tmk_core/protocol/usb_hid/USB_Host_Shield_2.0-tmk
new file mode 160000
index 00000000..48b2a095
--- /dev/null
+++ b/tmk_core/protocol/usb_hid/USB_Host_Shield_2.0-tmk
@@ -0,0 +1 @@
+Subproject commit 48b2a0950496c7b8ba1bb6ddb937b6f00a6de781